A new Stanford University study suggests small language models (SLMs) running on local PCs are matching or outperforming large language models (LLMs) in most tasks. The research compared QWEN 3, GEMMA 3, GPT-OSS, and GRANITE 4.0 on Nvidia and Apple M4 chips against cloud-based ChatGPT 5, Claude Sonnet 4.5, and Gemini 2.5 Pro. In chat tasks, the best SLM matched or exceeded LLM performance in 98.6% of cases.

For reasoning tasks, SLMs achieved 62.5% parity. Combined, SLMs performed as well or better in 81.2% of real-world scenarios. SLMs also cost 50-85% less in energy and compute.

Success rates for SLMs improved from ~50% in 2023 to 99% on easy tasks by October 2025, though they still lag in complex level 5 reasoning (51.5%). The study estimates the US SLM market at $10 trillion, threatening hyperscaler data center investments and potentially reducing demand for advanced chips.